A thermal imaging image optimization system based on multi-mode noise reduction

Abstract

This invention discloses a thermal imaging image optimization system based on multi-mode denoising, belonging to the field of image enhancement technology. It includes a region segmentation module, a multi-mode denoising module, a parameter update module, and a quality assessment module. This invention matches time-domain, spatial-domain, and frequency-domain denoising modes according to the motion state, texture complexity, and frequency-domain noise distribution characteristics of different regions in the thermal imaging image, and dynamically adjusts the fusion relationship between the multi-domain denoising results. This suppresses random noise, fixed-mode noise, and high-frequency disturbance noise while preserving thermal edge structure, texture details, and the continuity of moving regions. It avoids edge blurring, motion blur, and false filtering of real structures that are easily encountered in traditional single-domain denoising processes, reduces image flicker caused by parameter abrupt changes, and improves the temporal stability, structural integrity, and overall image clarity during the thermal imaging image denoising process.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and in particular to a thermal imaging image optimization system based on multi-mode noise reduction. Background Technology

[0002] In the field of thermal imaging image optimization, various image quality improvement methods have been disclosed in existing technologies. Traditional thermal imaging denoising methods mostly focus on processing within a single domain: spatial domain methods include Gaussian filtering, median filtering, bilateral filtering, etc., which directly smooth pixel values; temporal domain methods include multi-frame averaging, temporal Kalman filtering, etc., which utilize the correlation between consecutive frames to suppress random noise. Among frequency domain methods, a data processing method and system for infrared thermal imaging first acquires infrared thermal imaging information, and then performs denoising operations based on this information to obtain denoised wavelet coefficient values; then, based on the denoised wavelet coefficient values, a threshold function is constructed to obtain a local weighted threshold function and a high threshold function; next, an image reconstruction operation is performed based on the denoised wavelet coefficient values, the local weighted threshold function, and the high threshold function to obtain reconstructed image data; a first image enhancement operation is performed based on the reconstructed image data to obtain a first enhanced image; a second image enhancement operation is performed based on the first enhanced image to obtain a second enhanced image; finally, an image output operation is performed based on the second enhanced image to obtain an output image. In addition, recent deep learning methods, such as a deep learning image denoising method for infrared thermal imaging systems, first acquire the target image collected by the infrared thermal imaging system and perform downsampling processing to segment the image into several sub-images; extract image blocks and reconstruct pixels for each sub-image to obtain a preprocessed sub-image; then construct a noise map and add it as an extra channel to the preprocessed sub-image to form the input image; feed the input image into a pre-trained deep learning model to output an estimate of the noise; finally, remove the noise estimate from the original target image to obtain the denoised image.

[0003] For example, Chinese invention patent CN119515722B discloses a data processing method and system for infrared thermal imaging, comprising: acquiring infrared thermal imaging information; performing denoising operations based on the infrared thermal imaging information to obtain denoised wavelet coefficient values; performing threshold function construction operations based on the denoised wavelet coefficient values ​​to obtain a local weighted threshold function and a high threshold function; performing image reconstruction operations based on the denoised wavelet coefficient values, the local weighted threshold function, and the high threshold function to obtain reconstructed image data; performing a first image enhancement operation based on the reconstructed image data to obtain a first enhanced image; performing a second image enhancement operation based on the first enhanced image to obtain a second enhanced image; and performing an image output operation based on the second enhanced image to obtain an output image.

[0004] For example, Chinese invention patent CN113034408B discloses an infrared thermal imaging deep learning image denoising method and apparatus, comprising: acquiring a target image collected by an infrared thermal imaging system and performing downsampling processing to segment it into sub-images; extracting image blocks and recombining pixels in each sub-image to obtain a preprocessed sub-image; constructing a noise map and adding the noise map as an extra channel to the preprocessed sub-image as an input image; inputting the image into a trained deep learning model and outputting an estimate of the noise; removing the noise estimate from the target image to obtain a denoised image.

[0005] The above-mentioned technology has at least the following technical problems: Noise in thermal imaging images exhibits different characteristics in the spatial, temporal, and frequency domains. Noise reduction in only a single domain will ignore the valuable information in other domains that distinguishes signals from noise. Spatial domain filtering is effective for random noise in flat areas, but it easily blurs edges and cannot distinguish between fixed-pattern noise and the true temperature signal. The temporal domain can suppress random noise and retain motion information by utilizing multi-frame correlation, but existing methods rarely incorporate temporal modeling or only perform simple inter-frame averaging, resulting in motion blur in moving areas. Frequency domain transformation can set the noise reduction intensity for different frequency bands to remove low-frequency strip noise, but it may filter out true edges. Existing technologies struggle to retain the details and edge information of thermal images while suppressing multi-source noise. Summary of the Invention

[0006] This invention provides a thermal imaging image optimization system based on multi-mode noise reduction, which includes: a region segmentation module, a multi-mode noise reduction module, a parameter update module, and a quality assessment module.

[0007] The region segmentation module is used to divide image pixels into static flat regions, static edge regions, and moving regions based on two consecutive thermal imaging images through motion detection and local texture analysis.

[0008] The multi-mode noise reduction module is used to assign exclusive noise reduction mode combinations to different regions and obtain the noise reduction mode parameters of the current frame. The stationary flat area uses the time domain noise reduction mode and the frequency domain noise reduction mode in combination. The stationary edge area uses the spatial domain noise reduction mode. The moving area uses the spatial domain noise reduction mode and the frequency domain noise reduction mode in combination. The noise reduction results of each region are filled with the original pixels and then gradually fused with boundary weights to obtain the initial noise-reduced image.

[0009] The parameter update module is used to adaptively adjust the noise reduction mode parameters of the time-domain noise reduction mode and the frequency-domain noise reduction mode based on the noise residual variance of the static flat area in the initial noise reduction image of the current frame.

[0010] The quality assessment module is used to perform a no-reference quality assessment on the initial denoised image and, based on the assessment results, decide whether to trigger the multi-mode denoising module to perform local iterative denoising on a specific area, ultimately outputting an optimized thermal imaging image.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention provides a thermal imaging image optimization system based on multi-mode denoising. By jointly analyzing the motion features, local texture features, and frequency domain noise distribution features in thermal imaging images, it calls time domain, spatial domain, and frequency domain denoising modes to perform differentiated collaborative denoising processing for stationary flat areas, stationary edge areas, and moving areas, respectively. This allows for adaptive matching of the optimal denoising strategy in different areas, thereby achieving joint suppression of random noise, fixed-mode noise, and high-frequency disturbance noise. At the same time, it enhances the ability to preserve thermal edge structure, the continuity of moving areas, and the overall image detail fidelity. This solves the problem that single-domain denoising in the prior art cannot simultaneously take into account the suppression of multi-source noise and the preservation of thermal image detail edges, as well as the problems that traditional time domain denoising is prone to motion blur, spatial domain denoising is prone to edge blurring, and frequency domain denoising is prone to misfiltering real structural information.

[0012] 2. This invention performs joint analysis of thermal imaging images by combining local variance and inter-frame average difference, achieving adaptive region division of static flat areas, static edge areas, and moving areas. It also dynamically allocates corresponding noise reduction modes according to the image characteristics of different areas, so that the time-domain, spatial-domain, and frequency-domain noise reduction methods can play a role in the most suitable areas. This achieves a synergistic balance between random noise suppression, thermal edge structure preservation, and motion region continuity enhancement, solving the problem that the unified noise reduction strategy in the prior art cannot simultaneously take into account the noise characteristics and structural features of different areas.

[0013] 3. This invention dynamically fuses and adjusts the spatial domain noise reduction results and frequency domain noise reduction results in the moving region based on local variance and the proportion of high-frequency noise energy. This enables the system to adaptively adjust the fusion relationship by simultaneously considering the complexity of the local structure and the high-frequency disturbance state in the frequency domain. This enhances the ability to preserve the spatial domain structure when the edge structure is obvious and the noise is weak, and improves the ability to suppress the frequency domain noise when the high-frequency noise is enhanced. In this way, a dynamic balance is achieved between the noise suppression ability and the edge continuity preservation ability in the moving region, solving the problem that the fixed fusion weights in the prior art are difficult to adapt to complex motion scenes and changes in multi-source high-frequency noise.

[0014] 4. This invention performs a no-reference quality assessment on the initial denoised image and recursively updates and locally iteratively enhances the denoising parameters based on the comprehensive quality assessment results. This enables the system to dynamically adjust the time-domain and frequency-domain denoising parameters according to the current denoising effect, solving the problems of fixed denoising parameters, insufficient scene adaptability, and poor denoising stability in complex noise environments in the prior art. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A structural diagram of a thermal imaging image optimization system based on multi-mode noise reduction provided in this application embodiment; Figure 2 This is a flowchart of the noise reduction steps involved in this embodiment; Figure 3 The image shown is a result of the thermal imaging region division in this embodiment, where (a) represents the original thermal imaging image and (b) represents the result of the region division. Figure 4 This is a flowchart illustrating the time-domain and frequency-domain collaborative noise reduction process involved in this embodiment; Figure 5 The above are comparison diagrams of the frequency domain energy distribution of the thermal imaging image before and after denoising in this embodiment. Figures (c) and (d) represent the frequency domain energy spectra of the original thermal imaging image and the multi-mode denoising optimized image, respectively. Figure 6 This is a graph showing the change in the comprehensive quality evaluation value involved in this embodiment; Figure 7 The figures shown are comparisons of the regional detail preservation and noise reduction effects involved in this embodiment. Figure (e) represents the original thermal imaging image, Figure (f) represents the multi-mode noise reduction optimized image, and Figures (g) and (h) represent magnified images of local details of the original thermal imaging image and the multi-mode noise reduction optimized image, respectively. Detailed Implementation

[0017] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0018] Embodiment 1 of the present invention Figure 1 The diagram shown is a structural diagram of a thermal imaging image optimization system based on multi-mode denoising provided in an embodiment of this application. The system includes: a region division module, a multi-mode denoising module, a parameter update module, and a quality assessment module.

[0019] The region segmentation module is used to divide image pixels into static flat regions, static edge regions, and moving regions based on two consecutive thermal imaging images through motion detection and local texture analysis.

[0020] like Figure 2 The flowchart of the denoising steps involved in this embodiment includes: jointly analyzing the motion change features and local texture features in continuous thermal imaging images to achieve adaptive segmentation of static flat regions, static edge regions, and moving regions; applying time-domain, spatial-domain, and frequency-domain denoising methods to perform differentiated collaborative denoising processing for different regions; generating an initial denoised image through gradual fusion of region boundaries; adaptively updating denoising parameters based on the noise residual state in the denoising result; and finally comprehensively evaluating the noise suppression capability and structure preservation capability in the denoising result through a no-reference quality assessment, and performing iterative enhancement processing on regions with insufficient local quality, thereby achieving multi-mode collaborative adaptive denoising and image quality optimization of thermal imaging images in complex scenes.

[0021] The system receives the current frame and the previous frame of thermal imaging images. It accumulates the pixel grayscale values ​​in the current frame row by row to generate an integral image, and simultaneously accumulates the squares of the pixel grayscale values ​​row by row to generate a square integral image. The integral image records the cumulative sum of all pixel grayscale values ​​from the top left corner of the image to the current position, while the square integral image records the cumulative sum of the squared grayscale values ​​of pixels within the corresponding region. A pixel analysis window of a preset size is established centered on the pixel to be analyzed. Based on the integral image and the square integral image, the system quickly obtains the pixel analysis results through addition and subtraction operations at the four vertices of the window. The sum of gray levels and the sum of squared gray levels within the pixel analysis window are obtained quickly by adding and subtracting the integral image values ​​at the four corner points using the inclusion-exclusion principle. Specifically, the influence of overlapping areas is eliminated by subtracting the upper right and lower left corners from the lower right corner and then adding the upper left corner. Subsequently, the obtained sum of gray levels and sum of squared gray levels are divided by the window area to calculate the local mean and mean square value. Finally, according to the definition of variance, the mean square value is subtracted from the mean square value to obtain the local variance, thereby realizing the rapid analysis of the local gray level fluctuation characteristics of each pixel in the thermal imaging image.

[0022] The inclusion-exclusion principle is a method for calculating the deduplication count of overlapping regions. Its core idea is to obtain an accurate result by first calculating the whole region, then gradually subtracting the duplicate parts, and finally replenishing the parts that were subtracted. In the problem of summing a two-dimensional rectangular region, the integral image is first used to obtain the cumulative sum of the large region from the origin to the lower right corner. Then, the sums of the regions contained in the upper right and lower left corners are subtracted respectively. However, since the overlapping regions of these two parts are subtracted once, the overlapping part corresponding to the upper left corner needs to be added back. Thus, the true cumulative value inside the target window is accurately obtained through addition and subtraction compensation. Essentially, it is to eliminate the overlapping contribution by block superposition and cancellation, so as to achieve efficient solution of complex regions.

[0023] If the local variance of a pixel is less than or equal to the preset flatness threshold, it indicates that the gray-level variation in the area surrounding the pixel is small, and there is a high gray-level consistency between adjacent pixels in the area. This usually corresponds to a region with a stable temperature distribution in a thermal imaging image, so it is marked as a flatness candidate point. If the local variance of a pixel is greater than the preset flatness threshold, it indicates that there is a significant gray-level variation in the area surrounding the pixel, and there is a large temperature difference between adjacent pixels. This usually corresponds to the edge or texture structure of a thermal target. This type of region contains more real edge details and high-frequency structural information, so it is marked as an edge candidate point so that a noise reduction mode with stronger edge preservation capabilities can be used to perform differentiated processing on the corresponding region.

[0024] The flatness threshold in the database is set by professional technicians based on the statistical results of local gray-level variance in historical thermal imaging samples. By establishing local analysis windows on a large number of labeled thermal imaging images, the local gray-level variance distribution ranges in flat background areas and edge structure areas are statistically analyzed separately. The flatness threshold is determined based on the intersection interval of the local gray-level variance distributions of the two types of areas. Therefore, professional technicians write the local gray-level variance critical value that can effectively distinguish between flat areas and edge areas into the database as the preset flatness threshold. At the same time, it allows professional technicians to adjust according to different scenarios to improve the accuracy of area division in different application scenarios.

[0025] In this embodiment, the preset logic for the pixel analysis window size, motion threshold, local analysis window size, edge threshold, local frequency domain analysis window size, boundary transition band width, local statistical window size, upper limit of noise residual variance, smoothing weight, fusion evaluation coefficient, quality threshold, iteration number threshold, and flatness threshold in the database is similar. The specific values ​​can be flexibly set by professional technicians in combination with business scenarios, equipment characteristics, and control requirements. This embodiment will not be elaborated further.

[0026] By performing a difference operation on the pixel grayscale values ​​at corresponding positions in the previous and current thermal imaging images, an inter-frame difference map is generated. The pixel values ​​in the inter-frame difference map represent the degree of grayscale change at corresponding positions between two consecutive frames. Since the temperature distribution of stationary areas usually changes little between two consecutive frames, while the grayscale values ​​at corresponding positions in moving target areas will change significantly due to target displacement, attitude changes, or thermal distribution changes, the inter-frame difference value can be used to characterize the motion change features in thermal imaging images. The pixel values ​​in the inter-frame difference map are accumulated row by row to generate a difference integral map. A pixel analysis window of a preset size is established with the pixel to be analyzed as the center. Based on the difference integral map, the sum of differences within the pixel analysis window is quickly obtained through the inclusion-exclusion principle, i.e., the addition and subtraction operations at the four vertices of the window. Then, the average difference value of the corresponding pixel is calculated based on the sum of differences and the number of pixels in the window. Using the local window average difference instead of the difference value of a single pixel can reduce the impact of random thermal noise on motion detection results, avoid single-point noise fluctuations being misjudged as moving areas, and thus improve the stability and robustness of motion area recognition.

[0027] When the average difference value is greater than or equal to the preset motion threshold, it indicates that there is a significant grayscale change between two consecutive thermal imaging frames of the current pixel, and the temperature distribution in the region has undergone a strong temporal change. This usually corresponds to a moving target and a target displacement region, so the corresponding pixel is marked as a motion candidate point. When the average difference value is less than the preset motion threshold, it indicates that the grayscale change between two consecutive frames of the current pixel is small, and the temperature distribution in the region remains generally stable. This usually corresponds to a static background region or a non-moving region, so no motion candidate marking is performed on the corresponding pixel. By performing motion candidate determination based on the local average difference value rather than the instantaneous difference value of a single pixel, the interference of random thermal noise, local flicker, and isolated abnormal pixels on the motion detection results can be reduced, thereby improving the stability and accuracy of motion region recognition.

[0028] Based on the labeling results of flat candidate points, edge candidate points, and motion candidate points, the thermal imaging image is classified into regions. Pixel sets with motion candidate point labels are classified as motion regions because these regions exhibit significant temporal variations between consecutive thermal imaging frames. Continuing to use strong temporal fusion in these regions could easily produce motion blur, thus requiring a separate noise reduction mode suitable for moving targets. For pixel sets without motion candidate point labels but with edge candidate point labels, the system classifies them as stationary edge regions. This is because while these regions do not exhibit significant motion changes, their local grayscale variance is large, typically containing information about thermal target edges, texture structures, or temperature abrupt changes. Therefore, a noise reduction method with strong edge preservation capabilities is needed to avoid excessive smoothing of edge details. For pixel sets without motion candidate point labels but with flat candidate point labels, the system classifies them as stationary flat regions. This is because these regions exhibit high consistency in both the spatial and temporal domains, with fewer real high-frequency structures, making them more suitable for using a combined temporal and frequency domain approach to enhance random noise suppression, thus providing a region-level processing basis for subsequent multi-mode collaborative noise reduction.

[0029] like Figure 3 The figure shown is a result of the thermal imaging image region division involved in this embodiment. (a) represents the original thermal imaging image, and (b) represents the region division result. Different regions in the thermal imaging image are visualized. The stationary flat area is marked in blue, the stationary edge area is marked in yellow, and the moving area is marked in red.

[0030] After dividing the region into moving, stationary edge, and stationary flat regions, corresponding denoising processing combinations were assigned to each region based on their spatial and temporal grayscale variation characteristics. Specifically, the stationary flat region exhibits relatively small local grayscale variations and no significant motion changes, with limited real edge and texture information, making it more suitable for a combined temporal and frequency domain denoising approach to enhance random noise suppression. Conversely, the stationary edge region shows larger local grayscale variations but no significant motion changes, containing more thermal target edge and texture structure information; therefore, the spatial domain denoising approach, with its stronger edge-preserving capabilities, is prioritized. Noise reduction mode: Pixels in the moving region exhibit both temporal and local structural variation features. To avoid motion blur caused by temporal fusion, and to balance edge continuity and random noise suppression in moving targets, the system employs a combined spatial and frequency domain noise reduction mode for noise reduction processing. Subsequently, based on the noise reduction mode corresponding to the current region, the system reads the corresponding noise reduction mode parameters from the database. These parameters include a temporal weighting factor, a frequency threshold coefficient, a spatial regularization parameter, and a fusion ratio parameter between different modes. This allows the current frame of the thermal imaging image to perform multi-mode collaborative noise reduction processing on the corresponding region.

[0031] The multi-mode noise reduction module is used to assign exclusive noise reduction mode combinations to different regions and obtain the noise reduction mode parameters of the current frame. The stationary flat area uses the time domain noise reduction mode and the frequency domain noise reduction mode in combination. The stationary edge area uses the spatial domain noise reduction mode. The moving area uses the spatial domain noise reduction mode and the frequency domain noise reduction mode in combination. The noise reduction results of each region are filled with the original pixels and then gradually fused with boundary weights to obtain the initial noise-reduced image.

[0032] like Figure 4 The flowchart shown is a time-domain and frequency-domain collaborative noise reduction process involved in this embodiment. It includes: for stationary flat regions, time-domain noise reduction and frequency-domain noise reduction are used in a collaborative process. The time-domain noise reduction is performed by performing a recursive weighted average operation on the corresponding pixels in the current frame and the previous frame of the thermal imaging image. At the same time, the frequency-domain noise reduction is performed by performing wavelet decomposition on the current frame of the thermal imaging image to divide the image into low-frequency structural information and high-frequency detail information. The soft threshold is adaptively calculated based on the noise intensity in the high-frequency detail sub-band. The high-frequency wavelet coefficients are subjected to soft threshold attenuation processing to suppress high-frequency random noise and retain the main structural information. After completing the inverse wavelet reconstruction, the time-domain noise reduction result and the frequency-domain noise reduction result are fused at the corresponding positions in the stationary flat region to obtain the joint noise reduction result of the stationary flat region.

[0033] For stationary flat regions, the temporal domain noise reduction mode is invoked. Pixel values ​​at corresponding positions within the stationary flat regions are read from the current and previous thermal imaging images. A recursive weighted average is then performed on the corresponding pixels. Since pixels in stationary flat regions typically do not exhibit significant motion changes between consecutive frames, the actual temperature distribution within the region shows high temporal consistency. In contrast, random thermal noise usually fluctuates randomly between different frames. Therefore, performing temporal domain linear fusion on two consecutive thermal imaging images can enhance stable temperature information and reduce random noise fluctuations. Based on a preset temporal domain weighting factor, the current... The pixel values ​​of the previous frame are weighted and fused with the corresponding pixel values ​​of the previous frame. The time-domain weighting factor is used to control the proportion of thermal imaging information from the previous frame in the fusion result. When the time-domain weighting factor increases, the proportion of information from the previous frame increases, and the random noise averaging effect is enhanced. When the time-domain weighting factor decreases, the proportion of information from the current frame increases. The time-domain fusion result value is the sum of one minus the time-domain weighting factor multiplied by the time-domain weighting factor and the current frame pixel value multiplied by the corresponding pixel value of the previous frame. After completing the recursive weighted averaging operation of all pixels in the static flat area, a time-domain denoising result map is generated based on the time-domain fusion result value of each pixel.

[0034] The frequency domain noise reduction mode performs wavelet decomposition processing on the current frame of the thermal imaging image. It employs a two-dimensional discrete wavelet transform to perform multi-level wavelet decomposition processing on the thermal imaging image, where the wavelet basis functions use the Daubechies wavelet basis. Wavelet (Dobessie wavelet) decomposition layers are preset by professionals to be two to four layers based on the resolution of the thermal imaging image and the noise frequency distribution range. In each layer of wavelet decomposition, convolution operations are performed on the row and column directions of the thermal imaging image through low-pass and high-pass filters, respectively, and combined with downsampling to obtain a low-frequency approximate subband and three high-frequency detail subbands. The low-frequency approximate subband is used to characterize the main temperature distribution and large-scale thermal structure information in the thermal imaging image, and the three high-frequency detail subbands are used to characterize horizontal edges, vertical edges, and diagonal textures, respectively. After completing the multi-layer wavelet decomposition, the main structure information in each low-frequency approximate subband is retained, and soft threshold attenuation processing is performed on the wavelet coefficients in each high-frequency detail subband to suppress high-frequency random noise. After completing the wavelet coefficient processing of all high-frequency subbands, inverse transform operation is performed according to the inverse wavelet reconstruction order corresponding to the decomposition process to obtain the frequency domain denoising result image, thereby realizing the separation and joint optimization of high-frequency random noise and real thermal structure information in the frequency domain in the thermal imaging image.

[0035] Statistical analysis is performed on the wavelet coefficients in the high-frequency detail subband, and the set of absolute values ​​of the coefficients in the high-frequency detail subband is selected. The dispersion of these coefficients is statistically analyzed to reflect the noise intensity. The median is calculated after taking the absolute value of the coefficients in the subband, and the absolute deviation of the median is used to approximate the noise level, thereby obtaining an estimate of the noise standard deviation. Since the random thermal noise in thermal imaging images mainly manifests as high-frequency random fluctuations, the noise standard deviation can be used to characterize the noise intensity in the current high-frequency detail subband. Subsequently, the corresponding soft threshold is calculated by multiplying the noise standard deviation by a preset threshold coefficient. The larger the noise standard deviation, the stronger the random noise in the current high-frequency detail subband, and the larger the corresponding soft threshold is.

[0036] If the absolute value of the wavelet coefficient corresponding to a certain pixel is less than or equal to the soft threshold, it indicates that the high-frequency energy corresponding to the wavelet coefficient is weak. Its variation amplitude mainly comes from random thermal noise fluctuations rather than the real edge or texture structure. Therefore, the corresponding wavelet coefficient is attenuated to zero to eliminate the weak high-frequency components generated by random thermal noise. By performing zeroing processing on wavelet coefficients below the soft threshold, the random high-frequency noise components in the thermal imaging image can be reduced, while avoiding a large number of weak noise signals being re-superimposed on the thermal imaging image during the inverse wavelet reconstruction process.

[0037] If the absolute value of the wavelet coefficient corresponding to a certain pixel is greater than the soft threshold, it indicates that the wavelet coefficient not only contains random thermal noise components, but also contains real high-frequency information such as the edge of the thermal target, texture structure, or temperature change. Therefore, the corresponding soft threshold is subtracted while retaining the original wavelet coefficient sign direction to weaken the random noise component and retain the main structural information. Retaining the original sign direction can maintain the consistency of the edge change direction and texture structure in the thermal imaging image, and subtracting the soft threshold can reduce the amplitude of high-frequency random fluctuations. After completing the soft threshold attenuation processing of the wavelet coefficients of all high-frequency detail subbands, the inverse wavelet transform is performed by combining the retained low-frequency subbands and the processed high-frequency detail subbands to generate a frequency domain noise reduction result map, thereby reducing random thermal noise while retaining the main structure and edge detail information in the thermal imaging image.

[0038] The pixel values ​​at corresponding positions in the static flat areas of the temporal and frequency domain denoising results are read, and mean fusion processing is performed on the corresponding pixels. Since the temporal denoising result mainly utilizes the temporal consistency between consecutive frames to weaken random thermal noise, it can effectively reduce temporal random fluctuations in the flat area. The frequency domain denoising result mainly utilizes soft threshold attenuation processing in the high-frequency detail subband to suppress high-frequency random noise, which can further weaken spatial high-frequency noise components. Therefore, the two denoising results correspond to different noise suppression mechanisms. By performing mean fusion on the temporal and frequency domain denoising results, the static flat area can simultaneously possess the ability to suppress temporal random noise and frequency high-frequency noise, thereby enhancing the overall denoising effect in the static flat area. After completing the fusion processing of all pixels in the static flat area, the fused pixel values ​​are written to the corresponding positions in the temporary flat area result.

[0039] For stationary edge regions, the spatial domain denoising mode is invoked to perform edge-preserving denoising on the current frame of the thermal imaging image. The spatial domain denoising mode employs a guided filtering algorithm, using the current frame of the thermal imaging image as the guide image. A local analysis window of a preset size is established around the pixel to be processed, and the gray values ​​of the pixels within the local analysis window are statistically analyzed to obtain the gray mean and gray variance of the corresponding window. The gray mean is used to characterize the overall brightness distribution characteristics of the current local region, and the gray variance is used to characterize the degree of gray change in the current local region. Since stationary edge regions usually contain thermal target edges, texture structures, and temperature change information, the gray changes between adjacent pixels are more obvious. Therefore, the gray statistical results in the local analysis window are used to analyze the edge intensity and structural change characteristics of the current region, thereby providing a local structural basis for edge preservation and smoothing intensity adjustment in subsequent guided filtering.

[0040] For a local analysis window where the local gray-level variance is less than the edge threshold, it indicates that the gray-level variation between adjacent pixels in the current region is small, there are few real edges and texture structures in the region, and the local gray-level fluctuations are more due to random thermal noise. Therefore, the local gray-level deviation value is obtained by subtracting the local gray-level variance from the edge threshold. The regularization parameter of the guided filter is increased based on the local gray-level deviation value to enhance the smooth constraint ability between adjacent pixels in the local analysis window, thereby improving the suppression strength of random thermal noise.

[0041] To achieve an adaptive control relationship between local grayscale deviation values ​​and the increase in the guided filter regularization parameter, a monotonic mapping function is constructed and stored in a database. First, the optimal regularization parameter corresponding to different local grayscale deviation intervals is statistically analyzed on the sample dataset. Professional technicians generate a monotonically increasing mapping function through fitting, such as using a piecewise linear function or a monotonic spline function, ensuring that the increase in the regularization parameter maintains a consistent trend with the increase in the local grayscale deviation value. Then, the mapping function is discretized into several key control points, forming a table structure of local grayscale deviation values ​​and the increase in the regularization parameter, which is then stored in the database. During actual operation, by querying the interval containing the current local grayscale deviation value, the corresponding control point in the database is retrieved and interpolated to obtain the increase in the regularization parameter in real time. This enables the mapping relationship to be updated, expanded, and quickly retrieved. After extracting the increase in the regularization parameter, it is added to the current regularization parameter to obtain the updated regularization parameter, which is then applied to subsequent pixels.

[0042] For local analysis windows where the local gray-level variance is greater than or equal to the edge threshold, it indicates that there is a significant gray-level change in the current region. The local region contains information about the edge of the thermal target, texture structure, or temperature change. Therefore, the local gray-level variance is subtracted from the edge threshold to obtain the local variance deviation value. Based on the local variance deviation value, the reduction value of the regularization parameter of the guided filter is determined to weaken the smoothing effect across the edge direction, thereby preserving the real edge structure and texture details in the thermal imaging image. After completing the guided filtering processing for all pixel positions, the system generates a spatial domain noise reduction result map and writes the pixel values ​​of the corresponding positions of the static edge area into the temporary edge result map.

[0043] To achieve a monotonically decreasing relationship between the regularization parameter and the guided filter based on the local variance deviation value, the regularization parameter corresponding to the optimal guided filter effect in different deviation intervals is first statistically analyzed based on training samples or historical images. Professional technicians then construct a mapping function that satisfies the monotonically decreasing constraint through regression fitting or piecewise function approximation, ensuring that the regularization parameter monotonically decreases as the deviation value increases. This continuous mapping function is then discretized, and the corresponding regularization parameter values ​​are extracted at key deviation nodes, forming a mapping table between the local variance deviation value and the regularization parameter decrease value, which is stored in a database as structured key-value pairs. In practical applications, the mapping points of adjacent intervals in the database are queried based on the real-time calculated local variance deviation value, and the current regularization parameter decrease value is obtained through linear interpolation. This enables the storage, retrieval, and real-time adaptive updating of this monotonically decreasing mapping relationship. After extracting the regularization parameter decrease value, the current regularization parameter is subtracted from the decrease value to obtain the updated regularization parameter, which is then applied to subsequent pixels.

[0044] For the motion region, both spatial domain denoising and frequency domain denoising modes are invoked simultaneously to perform joint denoising processing. Spatial domain denoising and frequency domain denoising are performed on the corresponding pixels of the motion region respectively, resulting in spatial domain denoising result images and frequency domain denoising result images. Then, a preset fusion ratio parameter between the spatial domain denoising result and the frequency domain denoising result is read from the database. The fusion ratio parameter is used to control the contribution of different denoising modes to the final motion region fusion result. At the same time, the local variance value calculated by the corresponding pixels of the motion region during the region segmentation stage is read. Since the local variance can characterize the intensity of local gray-level changes in the motion region, the fusion relationship between the spatial domain denoising result and the frequency domain denoising result is dynamically adjusted based on the local variance value and the frequency domain noise characteristics in the motion region, thereby achieving collaborative optimization of edge preservation and random noise suppression in the motion region.

[0045] A local frequency domain analysis window of a preset size is established centered on the corresponding position of a pixel in a certain motion region within the high-frequency detail sub-band. Energy statistical analysis is performed on the wavelet coefficients within the local frequency domain analysis window. Wavelet coefficients with absolute values ​​less than or equal to the soft threshold typically correspond to weak high-frequency components generated by high-frequency random thermal noise. Therefore, the sum of squares of wavelet coefficients with absolute values ​​less than or equal to the soft threshold within the local frequency domain analysis window is calculated as the noise energy within the window. Simultaneously, the sum of squares of wavelet coefficients from all high-frequency detail sub-bands within the local frequency domain analysis window is calculated as the total high-frequency energy within the current local frequency domain analysis window. The noise energy is then divided by the total high-frequency energy to obtain the proportion of high-frequency noise energy for the corresponding pixel in the motion region. A higher proportion of high-frequency noise energy indicates that the high-frequency components in the current motion region originate more from random thermal noise, while a lower proportion indicates that the high-frequency components in the current region originate more from real edges, texture structures, or thermal target contours. This provides a basis for the local frequency domain noise characteristics for subsequent dynamic adjustment of the fusion relationship between spatial domain denoising and frequency domain denoising results.

[0046] If the local variance of a pixel in a certain motion region is greater than or equal to the edge threshold and the proportion of high-frequency noise energy is less than the preset frequency band threshold, it indicates that there is a significant grayscale change in the current motion region. The local region may contain thermal target edges, texture structures, or temperature change information. Moreover, the high-frequency components in the current local frequency domain mainly come from real edges and texture structures, rather than random thermal noise. Therefore, the fusion ratio of the spatial domain denoising result is increased based on the deviation between the proportion of high-frequency noise energy and the frequency band threshold. The larger the deviation between the proportion of high-frequency noise energy and the frequency band threshold, the higher the proportion of real edge components in the current high-frequency structure. Correspondingly, the contribution ratio of the spatial domain denoising result in the final fusion result is increased to enhance the continuity of moving target edges and the ability to preserve texture details, and to reduce the detail weakening effect of frequency domain soft thresholding on real high-frequency structures.

[0047] The specific logic for improving the fusion ratio of spatial domain denoising results based on the deviation between the high-frequency noise energy ratio and the frequency band threshold is as follows: First, based on training samples or historical image data, under different noise levels and spectral distribution conditions, professional technicians calculate the energy ratio deviation by subtracting the high-frequency subband energy ratio from the frequency band threshold. The fusion ratio increment corresponding to the highest peak signal-to-noise ratio of the spatial domain denoising result is used as the supervision target. On this basis, professional technicians construct a monotonically increasing mapping function between the energy ratio deviation and the fusion ratio increment using a regression fitting approximation method. This ensures that when the high-frequency noise energy ratio is less than the frequency band threshold and the energy ratio deviation increases, the fusion ratio increment of the spatial domain denoising result monotonically increases to enhance the noise suppression effect. Subsequently, this continuous mapping function is discretized, and the corresponding fusion weight values ​​are extracted at key deviation nodes to form a mapping table of energy ratio deviation and fusion ratio increment. This table is stored in the database in the form of structured key-value pairs, thereby realizing the learnable construction, storable retrieval, and online adaptive update of this mapping relationship. The updated spatial domain denoising result fusion ratio is obtained by adding the fusion ratio and fusion ratio increment of the current spatial domain denoising result in the database.

[0048] If the local variance of a pixel in a certain motion region is greater than or equal to the edge threshold and the proportion of high-frequency noise energy is greater than or equal to the preset frequency band threshold, it indicates that there is a significant grayscale change in the current motion region, the local region contains strong high-frequency structural information, and the high-frequency components in the current local frequency domain mainly come from random thermal noise. The proportion of high-frequency random fluctuations in the total high-frequency energy is relatively high. Therefore, the fusion ratio of the frequency domain noise reduction result is increased according to the deviation between the proportion of high-frequency noise energy and the frequency band threshold. Among them, the larger the deviation between the proportion of high-frequency noise energy and the frequency band threshold, the stronger the high-frequency random thermal noise in the current local region. The system correspondingly increases the contribution ratio of the frequency domain noise reduction result in the final fusion result to enhance the suppression capability of high-frequency random thermal noise and reduce the interference of random high-frequency fluctuations in the motion region on the thermal target contour and temperature distribution stability.

[0049] The specific logic for improving the fusion ratio of frequency domain noise reduction results based on the deviation between the high-frequency noise energy proportion and the frequency band threshold is as follows: First, based on training samples or historical image data, under different noise levels and spectral distribution conditions, professional technicians calculate the energy proportion deviation by subtracting the frequency band threshold from the high-frequency subband energy proportion. The fusion weight corresponding to the highest peak signal-to-noise ratio or optimal structural similarity in the frequency domain noise reduction result is then used as the supervision target. Based on this, a monotonic mapping function between the energy proportion deviation and the frequency domain fusion ratio increment is constructed through regression fitting or monotonic approximation methods, ensuring that when the high-frequency noise energy proportion is high... When the frequency band threshold and the energy ratio deviation value increase, the fusion ratio increment of the frequency domain noise reduction result monotonically increases to enhance the suppression capability of structural noise and high-frequency interference. Subsequently, the continuous mapping function is discretized, and the corresponding fusion weight value is extracted at the key deviation node to form a mapping table between the energy ratio deviation value and the frequency domain fusion ratio increment. This table is stored in the database in the form of structured key-value pairs, thereby realizing the learnable construction, storable retrieval and online adaptive update of the mapping relationship. The fusion ratio of the frequency domain noise reduction result and the fusion ratio increment of the current database are added together to obtain the updated fusion ratio of the frequency domain noise reduction result.

[0050] If the local variance of a pixel in a certain motion region is less than the edge threshold, it indicates that the local gray-level change in the current motion region is small, there is no obvious edge structure, texture details or temperature change information in the region, and the current high-frequency components are generally weak. Therefore, it is considered that there is no obvious priority retention requirement for the spatial domain denoising results and the frequency domain denoising results in the current region. The fusion ratio between the spatial domain denoising results and the frequency domain denoising results preset in the database remains unchanged. The two denoising results are weighted and fused according to the default fusion ratio to maintain the basic denoising capability of the motion region while avoiding brightness changes or texture discontinuities in local areas due to frequent adjustments of the fusion ratio.

[0051] Based on the updated fusion ratio of spatial domain denoising results and the updated fusion ratio of frequency domain denoising results, a weighted fusion operation is performed on the spatial domain denoising results and frequency domain denoising results at the corresponding positions in the motion region. The spatial domain denoising results are mainly used to maintain the continuity of the moving target edge and the stability of the local structure, while the frequency domain denoising results are mainly used to suppress high-frequency random thermal noise in the motion region. By dynamically controlling the contribution ratio of the two denoising results in the final fusion result, a balance is achieved between maintaining the moving target edge and suppressing high-frequency random noise, generating a temporary motion result map for subsequent regional result merging processing.

[0052] like Figure 5The figure shows a comparison of the frequency domain energy distribution of the thermal imaging image before and after denoising in this embodiment. Figures (c) and (d) represent the frequency domain energy spectra of the original thermal imaging image and the multi-mode denoising optimized image, respectively. The original image has obvious directional energy distribution and high-frequency dispersion characteristics in its frequency domain energy spectrum. After multi-mode denoising optimization, the discrete high-frequency energy in the frequency domain is weakened and the energy distribution is more concentrated. This indicates that while the system improves the visual effect in the spatial domain, it also suppresses high-frequency noise and directional interference in the frequency domain.

[0053] Region mapping is performed on all pixel locations in the thermal imaging image. The pixels are traversed. When a pixel location is in a stationary flat area, the pixel value of the same location in the temporary flat result image is read and written to the corresponding location in the initial denoising image. When the region label corresponding to a pixel location is a stationary edge area, the pixel value of the same location in the temporary edge result image is read and written to the corresponding location in the initial denoising image. When the region label corresponding to a pixel location is a moving area, the pixel value of the same location in the temporary motion result image is read and written to the corresponding location in the initial denoising image.

[0054] After completing the region mapping for all pixel locations, a boundary transition band of preset width is established at the intersection of different region labels. This boundary transition band covers the mode switching area between adjacent regions. Since different noise reduction modes and fusion strategies are used for the static flat region, static edge region, and moving region, directly performing a hard switch at the region boundary can easily lead to sudden brightness changes, texture discontinuities, or visual artifacts at the boundary. Therefore, for any pixel in the boundary transition band, the distance from that pixel to the boundaries of two adjacent regions is calculated, and the corresponding region's noise reduction result is calculated using an inverse distance ratio. The fusion weight, which is the ratio of the boundary distance of another adjacent region to the sum of the two boundary distances, determines the fusion weight of a certain region. This ensures that the closer a pixel is to a certain region, the higher the weight of the noise reduction result of that region in the final fusion result. The noise reduction results of each region are multiplied by their corresponding fusion weights and then weighted and summed to obtain the final fused pixel value in the boundary transition zone. This allows the pixel values ​​in the boundary transition zone to smoothly transition between the noise reduction results of different regions, thereby reducing the region switching traces between different noise reduction modes, improving the overall visual continuity and structural naturalness of the thermal imaging image, and finally generating the initial noise-reduced image.

[0055] The parameter update module is used to adaptively adjust the noise reduction mode parameters of the time-domain noise reduction mode and the frequency-domain noise reduction mode based on the noise residual variance of the static flat area in the initial noise reduction image of the current frame.

[0056] After generating the initial denoised image, a local statistical window of a preset size is established for the pixels corresponding to the static flat areas in the initial denoised image, and the gray values ​​of the pixels within the local statistical window are statistically analyzed. Since the actual temperature distribution in the static flat areas is usually relatively uniform, and there are no obvious edge structures and texture changes in the areas, the gray value fluctuations in the local areas mainly come from residual random thermal noise. The system first calculates the local mean of the pixel gray values ​​within the local statistical window, and then calculates the sum of squared deviations between each pixel gray value and the local mean to characterize the degree of gray value dispersion in the local areas. Subsequently, the residual variance of noise in the static flat areas is calculated based on the sum of squared deviations and the number of pixels in the local statistical window. The larger the residual variance of noise, the stronger the random thermal noise fluctuations in the current denoising result. The smaller the residual variance of noise, the more stable the thermal noise suppression effect in the current static flat areas, thus providing a quality feedback basis for subsequent dynamic adjustment of denoising parameters.

[0057] When the residual noise variance is greater than or equal to the preset high variance threshold, it indicates that there are still strong random thermal noise fluctuations in the static flat areas of the current initial denoising image, and the current temporal domain denoising intensity is insufficient. Therefore, the residual noise variance is subtracted from the high variance threshold to obtain the residual noise deviation value. Based on the residual noise deviation value, the temporal domain weighting factor is increased to improve the fusion ratio of the thermal imaging information of the previous frame in the temporal domain recursive weighted average, enhance the random noise averaging effect between consecutive frames, and further weaken the residual thermal noise in the static flat areas. At the same time, in order to avoid the temporal domain weighting factor being too large, which would cause the residual of the previous frame to be too strong or affect the response capability of subsequent dynamic regions, the system limits the updated temporal domain weighting factor to no more than the preset upper limit value.

[0058] The specific method for determining the increment value of the time-domain weighting factor based on the residual noise deviation value is as follows: First, based on training samples or historical video sequence data, the residual noise deviation value is calculated to characterize the current degree of insufficient time-domain noise suppression. On this basis, professional technicians use the increment of the time-domain weighting factor corresponding to the minimum noise level when the recursive weighted average in the time domain as the supervision target. A monotonic mapping relationship between the residual noise deviation value and the increment value of the time-domain weighting factor is constructed through regression fitting or monotonic function approximation methods, so that when the residual noise deviation value increases, the increment value of the time-domain weighting factor increases monotonically, thereby increasing the proportion of historical frames in the fusion result to enhance the time-domain average noise reduction effect. The continuous mapping function is discretized, and the corresponding increment value of the weighting factor is extracted at the key deviation nodes to form a mapping table between the residual noise deviation value and the increment value of the time-domain weighting factor. This table is stored in the database in the form of structured key-value pairs, thereby realizing the learnable construction, storage and retrieval, and online adaptive update of the mapping relationship.

[0059] When the residual variance of noise is less than or equal to the preset low variance threshold, it indicates that the random thermal noise in the stationary flat area of ​​the current initial denoising image has been sufficiently suppressed, and the current temporal denoising intensity is relatively high. Therefore, the residual variance of noise is subtracted from the low variance threshold to obtain the residual noise deviation value. The temporal weighting factor is reduced based on the residual noise deviation value to reduce the fusion ratio of the previous frame's thermal imaging information in the temporal recursive weighted average. By reducing the temporal weighting factor, the contribution ratio of the current frame's thermal imaging information in the temporal fusion result can be increased, thereby reducing the impact of the previous frame's residue on subsequent image updates and avoiding the phenomenon of trailing, response lag, or weakening of dynamic details in local areas due to excessively high temporal fusion intensity. At the same time, to avoid the temporal weighting factor being too small, which would significantly reduce the random noise averaging capability between consecutive frames, the system limits the updated temporal weighting factor to no less than the preset lower limit value.

[0060] The specific logic for determining the reduction value of the time-domain weighting factor based on the noise residual deviation value is as follows: First, based on training samples or historical video sequence data, under different motion change intensities and residual noise levels, the noise residual deviation value is analyzed to characterize the risk of dynamic distortion caused by excessive introduction of historical information in the time domain. The weighting factor adjustment amount corresponding to ensuring structural unambiguity and optimal temporal noise suppression effect is used as the supervision target. A monotonic mapping relationship between the noise residual deviation value and the reduction value of the time-domain weighting factor is constructed through regression fitting or monotonic function approximation methods, so that when the noise residual deviation value increases, the reduction value of the time-domain weighting factor increases monotonically, thereby reducing the proportion of historical frames in the fusion result to suppress temporal residue and ghosting effects. The continuous mapping function is discretized, and the corresponding reduction amount of the weighting factor is extracted at the key deviation nodes to form a mapping table of noise residual deviation value and time-domain weighting factor reduction value, which is stored in the database in the form of structured key-value pairs, thereby realizing the learnable construction, storage and retrieval, and online adaptive update of the mapping relationship.

[0061] When the residual variance of noise is greater than the low variance threshold and less than the high variance threshold, it indicates that the level of random thermal noise in the current initial denoised image is within the preset stable range. The current temporal denoising intensity can maintain a balance between the ability to suppress random thermal noise and the dynamic response capability. Therefore, the current temporal weighting factor is kept unchanged, and the proportion of the previous frame fusion in the temporal recursive weighted average is not adjusted. By maintaining the current temporal weighting factor when the residual variance of noise is within the stable range, the frequent fluctuations of temporal denoising parameters can be avoided, which can lead to unstable brightness, oscillation of denoising intensity, or local visual flickering between consecutive frames. This improves the stability and consistency of the continuous output process of thermal imaging images.

[0062] After the dynamic adjustment of the time-domain weighting factor is completed, the updated time-domain weighting factor is written to the database as the parameter state of the current frame for persistent storage, so that subsequent frame processing can be called and parameters can be inherited in a consistent manner. When performing time-domain denoising processing on the next frame of thermal imaging image, the time-domain weighting factor stored in the previous frame is read from the database and used as the initial parameter input for the recursive weighted average operation of the current frame. This realizes the continuous updating and adaptive evolution of the time-domain denoising parameters in the time series, ensuring the stable connection of denoising intensity and smoothing effect between consecutive frames.

[0063] The frequency domain mode employs a wavelet soft thresholding denoising method. Its core lies in using wavelet transform to decompose the current frame of thermal imaging into different frequency sub-bands and suppressing noise components in the high-frequency detail sub-bands. The wavelet soft threshold is calculated based on a threshold coefficient and the noise standard deviation analysis results in the high-frequency detail sub-bands. The noise standard deviation characterizes the intensity level of random thermal noise in the high-frequency sub-bands, while the threshold coefficient adjusts the balance between overall denoising intensity and detail preservation. By multiplying the threshold coefficient by the noise standard deviation, an adaptive soft threshold suitable for the high-frequency detail sub-bands of the current frame can be obtained. This allows the subsequent attenuation process of the wavelet coefficients to be dynamically adjusted according to changes in noise intensity, effectively suppressing high-frequency random noise while preserving edge and texture details in the thermal imaging image as much as possible.

[0064] After generating the initial denoised image, statistical analysis is performed on pixels in the static flat region. A local statistical window is established within this region, and the mean and deviation of the pixel grayscale values ​​within the window are calculated. Since the static flat region typically does not contain obvious edge and texture changes, its grayscale fluctuations mainly originate from residual random thermal noise. Therefore, this region can serve as a reference area for assessing the overall noise level. The system further calculates the sum of squares of the deviations between the pixel grayscale values ​​within the local statistical window and the local mean, and uses this to obtain the residual noise variance. Subsequently, the residual noise variances of multiple local windows within the static flat region are summarized or averaged to obtain the global noise variance, which is used to characterize the overall residual noise level of the current initial denoised image and serves as an important basis for subsequent adaptive adjustment of frequency domain parameters.

[0065] When the global noise variance exceeds the preset high variance threshold, it indicates that strong random thermal noise components still remain in the static flat areas of the current initial denoising image, and the existing wavelet soft thresholding denoising strength is insufficient. Therefore, the global noise variance is subtracted from the high variance threshold to obtain the global noise variance deviation value, and the threshold coefficient increase value is determined based on the global noise variance deviation value. By increasing the threshold coefficient, the overall level of the soft threshold calculated from the noise standard deviation can be improved, making it easier for wavelet coefficients with smaller amplitudes in the high-frequency detail subbands to be attenuated or suppressed, thereby enhancing the filtering ability of high-frequency random thermal noise and improving the overall frequency domain denoising strength.

[0066] The specific mapping logic for determining the threshold coefficient increment based on the global noise variance deviation value is as follows: First, the global noise variance deviation value is calculated based on training samples or historical image data to characterize the degree of deviation of the current overall noise intensity from the baseline level. On this basis, the corresponding threshold coefficient increment, under the condition of ensuring optimal noise suppression effect and minimal loss of detail, is used as the supervision target. Professional technicians construct a monotonic mapping relationship between the global noise variance deviation value and the threshold coefficient increment value through regression fitting or monotonic approximation methods, so that when the global noise variance deviation value increases, the threshold coefficient increment value monotonically increases, forming a mapping table between the global noise variance deviation value and the threshold coefficient increment value. This table is stored in the database in the form of structured key-value pairs, thereby realizing the learnable construction, storable retrieval, and online adaptive update of this mapping relationship.

[0067] When the global noise variance is lower than the preset low variance threshold, it indicates that the residual noise level in the stationary flat area of ​​the current initial denoising image is low, and the frequency domain has achieved a good noise suppression effect. Therefore, the global noise variance deviation value is obtained by subtracting the low variance threshold from the global noise variance deviation value, and the threshold coefficient reduction value is determined based on the global noise variance deviation value. By reducing the threshold coefficient, the soft threshold calculated by the noise standard deviation can be reduced accordingly, so that more wavelet coefficients with smaller amplitudes but which may contain detailed information can be retained, thereby reducing the excessive suppression of high-frequency details and improving the detail fidelity and edge clarity of thermal imaging images under low noise conditions.

[0068] The specific mapping logic for determining the threshold coefficient reduction value based on the global noise variance deviation value is as follows: First, the global noise variance deviation value is calculated based on training samples or historical image data to characterize the degree of deviation of the current overall noise level from the baseline state. The threshold coefficient reduction value corresponding to achieving the best balance between noise suppression and detail preservation is used as the supervision target. Professional technicians construct a monotonic mapping relationship between the global noise variance deviation value and the threshold coefficient reduction value through regression fitting or monotonic function approximation methods. This ensures that when the global noise variance deviation value increases, the threshold coefficient reduction value monotonically increases, thereby reducing the threshold coefficient to enhance the system's sensitivity to detail regions and avoid excessive smoothing of details due to an overly strong threshold. A mapping table between the global noise variance deviation value and the threshold coefficient reduction value is formed and stored in the database in the form of structured key-value pairs.

[0069] When the global noise variance is higher than or equal to the low variance threshold but does not exceed the high variance threshold, it indicates that the residual noise level in the current initial denoising image is within the preset stable range. There is no need to enhance the frequency domain denoising intensity or reduce the degree of detail preservation. Therefore, the threshold coefficient in the wavelet soft thresholding is kept unchanged at the default value to maintain the balance of the current frequency domain denoising strategy. The determined threshold coefficient is stored as the final frequency domain parameter of the current frame and uniformly applied to the soft threshold calculation process of all high-frequency detail subbands when performing wavelet transform on the next frame of thermal imaging image, thereby ensuring the stability and continuity of the frequency domain denoising parameters in the time series.

[0070] Each parameter update employs a recursive smoothing update method. It reads the adjusted parameter values ​​obtained from noise assessment or rule-based updates in the current frame, and simultaneously reads the old parameter values ​​actually used during denoising processing of the previous frame's thermal imaging image. These two values ​​are used as inputs for the recursive update. The adjusted parameter values ​​are multiplied by a preset smoothing weight, and the old parameter values ​​are multiplied by one minus the smoothing weight. The two are then added together to form the final updated parameters. The smoothing weight controls the fusion ratio between the current adjustment result and the historical parameter state, thus avoiding drastic changes caused by single-frame noise fluctuations or local anomaly estimations during the parameter update process. Through this weighted fusion mechanism, the parameter update result maintains the inheritance of historical stable states while responding to the noise change trend of the current frame, ultimately yielding new parameter values ​​for parameter calling and control in subsequent thermal imaging image denoising processing.

[0071] In this embodiment, parameter updates include updates to parameters such as threshold coefficient, time domain weighting factor, and regularization coefficient.

[0072] The new smoothing parameter values ​​obtained through recursive smoothing updates are written to the database as the final control parameters for the current frame and persistently stored. This database records the continuous state of parameter evolution across frames and serves as a parameter scheduling and historical state management unit, providing a traceable parameter baseline for subsequent frames. Then, when the corresponding denoising mode is executed on the next frame of the thermal imaging image, the system first reads the new smoothing parameter values ​​stored in the previous frame from the database and uses them as the initial input parameters for the current frame's denoising processing. These parameters control the relevant weights or threshold settings in the time, frequency, or spatial domain denoising process, thereby achieving cross-frame parameter inheritance and dynamic adaptive updates, ensuring the stability and consistency of the multi-frame continuous denoising process.

[0073] The quality assessment module is used to perform a no-reference quality assessment on the initial denoised image and, based on the assessment results, decide whether to trigger the multi-mode denoising module to perform local iterative denoising on a specific area, ultimately outputting an optimized thermal imaging image.

[0074] like Figure 6 The figure shown is a curve illustrating the change of the overall quality evaluation value in this embodiment, which shows the trend of the overall quality evaluation value of the thermal imaging image with the number of iterations during the iterative enhancement process.

[0075] After generating the initial denoised image, a no-reference quality assessment process is performed. A local statistical window is established for static, flat regions, and the dispersion of pixel grayscale values ​​within the window is calculated to extract the residual variance of noise. This residual variance characterizes the intensity level of residual random thermal noise in flat regions. The noise residual variance is then converted into a dimensionless noise suppression quality index using a min-max normalization method. The smaller the noise residual variance, the closer the corresponding noise suppression quality index is to 1. Subsequently, for static edge regions, the average gradient magnitude of the corresponding regions in the original and initial denoised images is calculated. The gradient preservation rate is obtained by dividing the average gradient after denoising by the average gradient before denoising, characterizing the ability of the denoising process to preserve edge structure and texture details. The gradient preservation rate is then converted into a dimensionless structure preservation quality index using a min-max normalization method, characterizing the ability of the denoising process to preserve edge detail information. Finally, a weighted fusion method is used to comprehensively map and calculate the two types of indices. Where A represents the comprehensive quality evaluation value, B represents the noise suppression quality index, C represents the structure preservation quality index, and α represents the preset fusion evaluation coefficient in the database.

[0076] When the overall quality evaluation value is less than or equal to the preset quality threshold, it indicates that the current initial denoising image has not achieved the expected balance between noise suppression and edge preservation. Therefore, the quality deviation value is obtained by subtracting the overall quality evaluation value from the quality threshold. Based on the quality deviation value, the smoothing weight is increased to accelerate the convergence speed of the parameters towards the target adjustment value, thereby enhancing the system's response to changes in the current noise state. This allows the updated parameters to be applied more quickly to the subsequent thermal imaging image denoising process, and the multi-mode denoising process is re-executed. However, the entire image is not recalculated; instead, only the corresponding problem areas where the overall quality evaluation value is lower than the preset quality threshold are located. Within these areas, the spatial domain, temporal domain, or frequency domain denoising modules are called again for local reprocessing to improve local quality. At the same time, an iteration number threshold is set as a constraint to limit the number of iterations for quality assessment, parameter adjustment, and local re-denoising, avoiding infinite iteration that would increase computational overhead or cause excessive image smoothing. This ensures that the overall processing efficiency is controlled while maintaining denoising quality.

[0077] The specific mapping logic for determining the smoothing weight increment based on the quality deviation value is as follows: First, based on training samples or historical thermal imaging sequence data, the quality deviation value is analyzed. Professional technicians use the smoothing weight increment corresponding to the system's optimal dynamic response state as the supervision target. Through regression fitting or monotonic function approximation methods, a monotonic mapping relationship between the quality deviation value and the smoothing weight increment value is constructed, so that when the quality deviation value increases, the smoothing weight increment value monotonically increases. Subsequently, this continuous mapping function is discretized, and the corresponding smoothing weight increment value is extracted at the key quality deviation nodes to form a mapping table of quality deviation value and smoothing weight increment value. This table is stored in the database in the form of structured key-value pairs, thereby realizing the learnable construction, storable retrieval, and online adaptive update of this mapping relationship.

[0078] like Figure 7 The figures shown are comparisons of the regional detail preservation and noise reduction effects involved in this embodiment. The original image and the multi-mode noise reduction optimized image are compared in both overall and local magnification. Figure (e) represents the original thermal imaging image, Figure (f) represents the multi-mode noise reduction optimized image, and Figures (g) and (h) represent magnified local detail images of the original thermal imaging image and the multi-mode noise reduction optimized image, respectively. In the local area of ​​the original image, vertical stripes are more obvious. After multi-mode noise reduction optimization, the stripe interference and background fine noise in the local area are significantly reduced. While the vertical stripe interference and background fine noise in the original image are effectively reduced, some real details are also weakened. This reflects the inherent trade-off between noise suppression and detail preservation in the process of thermal imaging noise reduction. It preserves the target edge and the main body of the thermal structure to the greatest extent, while significantly suppressing stripe interference and background fine noise.

[0079] When the comprehensive quality evaluation value is greater than the preset quality threshold, it indicates that the current initial denoising image has reached the preset balance requirement between the residual variance of noise and the edge gradient preservation rate. That is, the residual random thermal noise in the static flat area has been effectively suppressed, while the structural information in the static edge area has been well preserved. It is determined that the current multi-mode collaborative denoising result meets the quality constraint condition. Therefore, the subsequent iterative adjustment and local reprocessing process is stopped, and the current processing result is used as the final output result, outputting the optimized thermal imaging image.

[0080] In Embodiment 2 of the present invention, while remaining unchanged from Embodiment 1, the parameter recursive smooth update can also include: hierarchical transition update, specifically comprising: Read the adjusted parameter values ​​obtained in the current frame and the old parameter values ​​used when performing noise reduction processing on the previous frame of thermal imaging image, and subtract the old parameter values ​​from the adjusted parameter values ​​to obtain the parameter deviation value.

[0081] The absolute value of the parameter deviation is compared with a preset multi-level transition threshold range. The corresponding parameter transition level is determined according to the range in which the absolute value of the parameter deviation falls. Different transition levels correspond to different parameter transition step sizes. When the absolute value of the parameter deviation is large, a higher level parameter transition step size is used to improve the convergence speed of the parameter to the adjusted parameter value. When the absolute value of the parameter deviation is small, a lower level parameter transition step size is used to reduce the parameter change amplitude and maintain parameter stability. The direction of parameter adjustment is determined by the sign of the parameter deviation value, i.e., a positive sign indicates an increase and a negative sign indicates a decrease. After completing the parameter transition adjustment, a new transition parameter value is obtained.

[0082] The specific process for presetting multi-level transition threshold intervals involves: based on training samples and historical thermal imaging image data, statistically analyzing the distribution of parameter deviations between the adjusted parameter values ​​and the old parameter values ​​under different noise intensities, parameter variation amplitudes, and noise reduction quality states. Combined with the parameter adjustment amplitudes required for the system to achieve stable noise reduction under different parameter deviation conditions, key deviation nodes in the parameter transition process are determined. Based on the range of parameter deviation values, professional technicians classify the parameter deviation intervals into multiple transition threshold intervals, such as low-level, medium-level, and high-level intervals. Lower-level intervals correspond to smaller parameter transition step sizes to maintain parameter stability, while higher-level intervals correspond to larger parameter transition step sizes to improve the system's rapid response capability to abnormal noise states. After completing the interval division, each transition threshold interval and its corresponding parameter transition level are stored in the database in the form of a structured mapping table.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-modal denoising based thermal imaging image optimization system, characterized in that, The system includes: a region segmentation module, a multi-mode noise reduction module, a parameter update module, and a quality assessment module; The region segmentation module is used to divide image pixels into static flat regions, static edge regions, and moving regions based on two consecutive frames of thermal imaging images through motion detection and local texture analysis. The multi-mode noise reduction module is used to assign exclusive noise reduction mode combinations to different regions and obtain the noise reduction mode parameters of the current frame. The stationary flat region uses the time domain noise reduction mode and the frequency domain noise reduction mode to reduce noise in combination. The stationary edge region uses the spatial domain noise reduction mode. The moving region uses the spatial domain noise reduction mode and the frequency domain noise reduction mode to reduce noise in combination. The noise reduction results of each region are filled with the original pixels and then gradually fused with the boundary weight to obtain the initial noise reduction image. The parameter update module is used to adaptively adjust the noise reduction mode parameters of the time domain noise reduction mode and the frequency domain noise reduction mode according to the noise residual variance of the static flat area in the initial noise reduction image of the current frame. The quality assessment module is used to perform a no-reference quality assessment on the initial denoised image, and decide whether to trigger the multi-mode denoising module to perform local iterative denoising on a specific area based on the assessment result, and finally output the optimized thermal imaging image.

2. The multi-modal denoising based thermal imaging image optimization system of claim 1, wherein: The method, based on two consecutive thermal imaging images, divides image pixels into static flat regions, static edge regions, and moving regions through motion detection and local texture analysis, specifically including: The system receives two consecutive thermal imaging images and uses an integral image to quickly calculate the local variance of each pixel within the pixel analysis window. Pixels with local variance less than or equal to a preset flatness threshold are marked as flatness candidate points, and pixels with local variance greater than the preset flatness threshold are marked as edge candidate points. An inter-frame difference map is constructed using the previous frame and the current frame, and an integral map is also used to calculate the average difference value of each pixel within the pixel analysis window; Pixels with an average difference greater than or equal to a preset motion threshold are marked as motion candidate points, and those with a difference less than or equal to the threshold are not marked. The set of pixels with motion candidate points is marked as the motion region; the set of pixels with edge candidate points but no motion candidate points is marked as the stationary edge region; and the set of pixels with flat candidate points but no motion candidate points is marked as the stationary flat region.

3. The multi-modal denoising based thermal image optimization system of claim 1, wherein: The method involves assigning specific noise reduction mode combinations to different regions and obtaining the noise reduction mode parameters of the current frame. For static, flat regions, a combination of time-domain and frequency-domain noise reduction modes is used for noise reduction. Specifically, this includes: Based on the image characteristics of each region, corresponding noise reduction processing combinations are assigned to different regions from spatial domain noise reduction mode, temporal domain noise reduction mode and frequency domain noise reduction mode, and the noise reduction mode parameters of the current frame are obtained from the database. For a static flat area, the time domain noise reduction mode is invoked, the current frame thermal imaging image and the previous frame thermal imaging image are read, a recursive weighted average operation is performed on the corresponding pixels in the static flat area, and the pixel values ​​of the current frame and the corresponding pixel values ​​of the previous frame are linearly fused according to the preset time domain weighting factor to obtain the time domain noise reduction result image. The frequency domain noise reduction mode performs wavelet decomposition on the current frame thermal imaging image, decomposing the thermal imaging image into a low-frequency sub-band and multiple high-frequency detail sub-bands. The low-frequency sub-band is used to preserve the main structure of the thermal imaging and the overall temperature distribution, while the high-frequency detail sub-band is used to characterize edge, texture and high-frequency random noise information. The corresponding soft threshold is calculated based on the noise standard deviation in the high-frequency detail subband, and soft threshold attenuation processing is performed on the wavelet coefficients in the high-frequency detail subband. If the absolute value of the wavelet coefficient of a certain pixel is less than or equal to the soft threshold, the corresponding wavelet coefficient will be attenuated to zero. If the absolute value of the wavelet coefficient of a certain pixel is greater than the soft threshold, the corresponding soft threshold is subtracted while retaining the original sign direction, and the frequency domain noise reduction result is obtained by reconstructing through inverse wavelet transform. The mean pixel values ​​at corresponding positions in the static flat area of ​​the time-domain denoising result map and the frequency-domain denoising result map are fused, and the fused result is written into the temporary flat result map.

4. The thermal imaging image optimization system based on multi-mode noise reduction as described in claim 1, characterized in that: The static edge region employs a spatial domain noise reduction mode, specifically including: For static edge regions, the system calls the spatial domain noise reduction mode. The spatial domain noise reduction mode uses guided filtering to process the current frame thermal imaging image. The current frame thermal imaging image is used as the guide image. A local analysis window of a preset size is established around the pixel to be processed, and the gray mean and gray variance of the pixels within the local analysis window are statistically analyzed. For windows where the local gray-level variance is less than the edge threshold, the regularization parameter of the guided filter is increased based on the deviation between the edge threshold and the local gray-level variance. For windows where the local gray-level variance is greater than or equal to the edge threshold, the regularization parameter of the guided filter is weakened based on the deviation between the local gray-level variance and the edge threshold. After filtering all pixels, a spatial domain noise reduction result map is generated, and the pixel values ​​at the corresponding positions of the static edge areas are written into the temporary edge result map.

5. The multi-modal denoising based thermal imaging image optimization system of claim 1, wherein: The motion region employs a combined spatial domain noise reduction and frequency domain noise reduction approach, specifically including: For the motion region, the spatial domain noise reduction mode and the frequency domain noise reduction mode are called simultaneously to perform joint noise reduction processing, and the spatial domain noise reduction result map and the frequency domain noise reduction result map are obtained respectively. The fusion ratio of the preset spatial domain noise reduction result and the frequency domain noise reduction result in the database is extracted, and the local variance value of the corresponding pixel in the motion region is read. Centered on a pixel in a certain motion region, a local frequency domain analysis window of a preset size is established in the high-frequency detail sub-band. The proportion of high-frequency noise energy of the pixel in the motion region is analyzed based on the sum of squares of wavelet coefficients whose absolute value is less than or equal to the soft threshold within the local frequency domain analysis window, and the sum of squares of wavelet coefficients of all high-frequency detail sub-bands within the local frequency domain analysis window. If the local variance of a pixel in a certain motion region is greater than or equal to the edge threshold and the proportion of high-frequency noise energy is less than the preset frequency band threshold, the fusion ratio of the spatial domain noise reduction result is increased according to the deviation between the proportion of high-frequency noise energy and the frequency band threshold. If the local variance of a pixel in a certain motion region is greater than or equal to the edge threshold and the proportion of high-frequency noise energy is greater than or equal to the preset frequency band threshold, then the fusion ratio of the frequency domain noise reduction result is increased according to the deviation between the proportion of high-frequency noise energy and the frequency band threshold. If the local variance of a pixel in a certain motion region is less than the edge threshold, then the fusion ratio of the spatial domain noise reduction result and the frequency domain noise reduction result is maintained. Based on the fusion ratio, the spatial domain denoising results and the frequency domain denoising results are weighted and fused to generate a temporary motion result map.

6. The multi-modal denoising based thermal imaging image optimization system of claim 1, wherein: The process of fusing the denoising results of each region with the original pixel fill and then gradually merging them through boundary weights to obtain the initial denoised image specifically includes: Traverse the pixels. When a pixel is in a static flat area, read the pixel value at the same position in the temporary flat result image and write it to the corresponding position in the initial denoising image. When the region label corresponding to a certain pixel position is a stationary edge region, the pixel value at the same position in the temporary edge result image is read and written to the corresponding position in the initial noise reduction image; When the region label corresponding to a certain pixel position is a motion region, the pixel value at the same position in the temporary motion result image is read and written to the corresponding position in the initial noise reduction image; After completing the region mapping of all pixel locations, a boundary transition band of preset width is established at the intersection of different regions. Weighted inter-region interpolation is performed on the pixels in the boundary transition band to generate the initial denoised image.

7. The multi-modal denoising based thermal imaging image optimization system of claim 1, wherein: The adaptive adjustment of the temporal denoising mode based on the residual variance of noise in the static flat region of the initial denoised image of the current frame specifically includes: After generating the initial denoised image, a local statistical window of a preset size is established for the pixels corresponding to the static flat area in the initial denoised image, and the sum of squared deviations between the pixel gray values ​​and the local mean within the local statistical window is calculated. The residual variance of noise in the static flat area is calculated based on the sum of squared deviations. When the residual variance of noise is greater than or equal to the preset high variance threshold, the time domain weighting factor is increased, and the updated time domain weighting factor is limited to not exceeding the preset upper limit. When the residual variance of noise is less than or equal to the preset low variance threshold, the time domain weighting factor is reduced, and the updated time domain weighting factor is limited to not being lower than the preset lower limit. When the residual variance of the noise is greater than the low variance threshold and less than the high variance threshold, the current time domain weighting factor remains unchanged. The updated temporal weighting factor is written to the database and called when temporal denoising is performed on the next frame of thermal imaging image.

8. The multi-modal denoising based thermal imaging image optimization system of claim 1, wherein: The noise reduction mode parameters of the frequency domain noise reduction mode specifically include: The frequency domain mode employs wavelet soft thresholding for denoising, which is based on the threshold coefficients and the noise standard deviation analysis in the high-frequency detail subbands. After generating the initial denoised image, the residual variance of noise is calculated from the static flat region of the initial denoised image and used as the global noise variance. When the global noise variance exceeds the high variance threshold, the threshold coefficient is increased; When the global noise variance is below the low variance threshold, reduce the threshold coefficient; The default value is maintained when the global noise variance is higher than or equal to the low variance threshold but does not exceed the high variance threshold. The adjusted threshold coefficients are used to perform wavelet transform on all high-frequency detail subbands of the next frame of thermal imaging image.

9. The thermal imaging image optimization system based on multi-mode denoising as described in claim 7 or 8 further includes recursive smooth parameter update, specifically including: Each parameter update uses a recursive smoothing update, which means reading the adjusted parameter value obtained in the current frame and the old parameter value used when performing noise reduction processing on the previous frame of thermal imaging image, and performing a weighted fusion operation on the adjusted parameter value and the old parameter value according to the preset smoothing weight to obtain the new smoothed parameter value; The new smoothing parameter values ​​are written to the database and then invoked when the corresponding noise reduction mode is executed in the next frame of the thermal imaging image.

10. The multi-modal denoising based thermal imaging image optimization system of claim 1, wherein: The process of performing a no-reference quality assessment on the initial denoised image and determining whether to trigger the multi-mode denoising module to perform local iterative denoising on a specific region based on the assessment results, ultimately outputting an optimized thermal imaging image, specifically includes: A no-reference quality assessment is performed on the initial denoised image. The noise residual variance in the static flat region is extracted. The gradient magnitudes of the corresponding positions before and after denoising are calculated for the pixels in the static edge region. The gradient preservation rate in the static edge region is analyzed. The comprehensive quality evaluation value is calculated based on the gradient preservation rate and the noise residual variance. When the overall quality evaluation value is less than or equal to the preset quality threshold, the smoothing weight of the adjusted parameter value is increased, and multi-mode noise reduction processing is re-executed. The number of quality iteration adjustments is limited to no more than the iteration number threshold, and each iteration only performs local re-noise reduction processing on the problem area corresponding to the overall quality evaluation value being lower than the preset quality threshold. When the overall quality evaluation value is greater than the preset quality threshold, the optimized thermal imaging image is output.

Images